A Startup Wants to Fight AI Bioweapons With AI

Frontier AI labs have repeatedly warned that their systems could lower the barrier to building dangerous bioweapons. On Wednesday, San Francisco startup Radical Numerics answered with a claim of its own: a new edge model called Omnii Defense that identifies both natural and AI-engineered pathogens using only a laptop and a USB gene sequencer. Because it runs without an internet connection, the company says it can be used by researchers working in the field on unreliable networks.

The company's cofounders come from the research lineage of Evo and Evo-2, open-source DNA foundation models developed during doctoral work at Stanford; they founded Radical Numerics in 2025. Cofounder and CEO Eric Nguyen says he expects an AI-driven arms race in biodefense analogous to cybersecurity, where attackers and defenders both lean on machine learning for advantage. "We felt that the biosecurity space was far behind on the defense side," he says.

Omnii Defense builds on Omnii, Radical's genome language model trained on biological information beyond DNA, including RNA, proteins, and epigenetic signals. The startup drew attention in October when it said it used Omnii to design custom bacteriophages — viruses that kill specific bacteria — against a drug-resistant strain of E. coli, and that the resulting cocktail treated a sick patient under special permission from the Food and Drug Administration. Tabula Bioscience, another AI biotech startup, was involved in that work but was not disclosed in Radical's announcement; George Church, a Harvard geneticist who advised both companies, resigned from his Radical role last week over the omission.

Technically, the model issues several predictions for a given sequence — the odds it is pathogenic, the odds it is AI-generated, and its closest taxonomic group — combined into a threat score from zero to one. That distinguishes it from earlier federally funded tools such as Finding Engineering-Linked Indicators (Felix), which compare sequences against libraries of known organisms and essentially search for cataloged matches, useful when a bad actor makes small tweaks to an existing pathogen but not designed for prediction. Learning functional patterns rather than exact matches could let AI models spot similarities even after genetic spelling changes, which Christine Parthemore, Radical's head of biodefense, calls a "step change." In one experiment, the team removed the Bundibugyo virus from training data first at the species level and then at the whole genus level; with only the species removed, the model still judged 33 percent of pushed Bundibugyo sequences pathogenic, and with the genus removed it flagged 22 percent, while non-AI tools gave essentially no indication of risk.

Notably, many scientists still consider AI-made bioweapons a low-likelihood threat, and the company's evidence so far consists of internal experiments rather than independent benchmarking. Bronwyn MacInnis of the Broad Institute of MIT and Harvard argues vigilance is warranted regardless: "If we can use technology to compete with it and hopefully get ahead of any kind of nefarious actors, that seems sensible." What to watch next is whether outside groups replicate the Bundibugyo-style holdout results, whether the threat score holds up on field-collected samples, and whether undisclosed partnerships like the Tabula collaboration become standard disclosure practice.